The 14 next big things in applied AI for 2025 - Fast Company - AI2Work Analysis

The past year has seen the first generation of truly capable edge AI laptops arrive on the market. The key enablers are mainstream mobile CPUs that now expose dedicated tensor cores, integrated GPUs that can handle moderate inference workloads, and a tightening regulatory environment that pushes hardware‑level safeguards into the procurement process. This article distills those developments into a set of facts, benchmarks, and concrete steps for decision makers who need to evaluate whether an on‑device AI strategy will deliver ROI.

CPU families that support tensor cores

• Intel Rocket Lake (13th Gen) : The i5‑13600H and i7‑13700H are the only mainstream mobile CPUs with a dedicated AI engine. Benchmarks from the TechBench AI Suite 2025 show ~75 GFLOPs for BERT‑like transformer inference at 90–95 W TDP, with a peak of ~85 GFLOPs under sustained workloads.

• AMD Ryzen PRO/EPYC AI line : The Ryzen PRO 7000 series (mobile) and EPYC 8003 (server‑grade) both ship a “Unified Compute Engine” that offers ~60–70 GFLOPs for similar models, but at lower power envelopes (45–55 W). The mobile variants are now available in laptops marketed as “AI‑Ready.”

Integrated GPU performance

• Intel Arc Mobile GPUs (e.g., Arc A770M) deliver ~30 % of a discrete RTX 4060’s throughput for vision inference at 30 fps on a 256‑token BERT model. This is sufficient for most enterprise use cases that involve moderate batch sizes.

• AMD RDNA 3 mobile GPUs (e.g., RX 680M) reach ~25 % of discrete RTX 4060 performance under the same test conditions, but benefit from lower latency in memory‑bound workloads.

Secure enclaves and TEE support

• Intel SGX is no longer supported on mobile silicon; instead, Intel’s “Intel Security Engine” (ISE) provides a lightweight TEE with ~12 % overhead for inference tasks that require model protection.

• AMD’s Secure Enclave offers similar isolation with an average 9–11 % performance penalty. Both are fully compliant with the emerging EU AI Act hardware safety and bias mitigation guidelines, though certification is still in its infancy.

Device

Model

AI Throughput (GFLOPs)

TDP (W)

Inference Latency (ms, 256‑token BERT)

Laptop A

Intel i5‑13600H + Arc A770M

75

95

28

Laptop B

AMD Ryzen PRO 7000 + RX 680M

65

55

32

Laptop C

Intel i7‑13700H + RTX 4060 (discrete)

110

140

18

These figures illustrate that a well‑balanced laptop with integrated GPU can match or exceed the performance of a mid‑range discrete GPU while consuming 30–40 % less power. For most enterprise workloads—real‑time translation, predictive maintenance, and lightweight computer vision—the difference between an “AI‑Ready” (integrated) device and an “AI‑Heavy” (discrete) device is marginal in terms of user experience but significant in cost and battery life.

The average price for an AI‑Ready laptop dropped from $3,800 to $3,400 between Q1 2024 and Q1 2025 thanks to improved manufacturing yields and a shift away from premium discrete GPUs. AI‑Heavy models remain in the $5,200–$6,500 range but have seen a 15 % price decline due to increased competition among OEMs.

From an enterprise perspective, the primary cost drivers are:

• Capital Expenditure : Initial purchase of AI‑Ready laptops is roughly $3,400 per unit. For a 100‑user deployment this equates to $340,000.

• Operational Savings : On‑device inference eliminates the need for high‑tier cloud subscriptions (average $350/month/user). Over three years that saves ~$1.26 million.

• Bandwidth & Data egress : Edge inference reduces data transfer by 80–90 %, translating to < $1,000 in monthly bandwidth savings at enterprise rates.

• Battery life and device turnover : AI‑Ready laptops now offer up to 15 % longer battery life under mixed workloads, reducing the frequency of replacements and associated logistics costs.

The EU AI Act’s hardware certification roadmap is still evolving. Current guidance from the European Commission recommends that manufacturers provide:

• Hardware‑level bias mitigation logs (e.g., model audit trails embedded in firmware).

• Secure boot and signed firmware updates.

• TEE support for proprietary model protection.

In practice, enterprises can mitigate the risk of model theft by deploying Intel ISE or AMD Secure Enclave on all edge devices. The performance impact is modest (≈10 %) and far outweighed by the regulatory and IP benefits. Firmware tampering risks are addressed through signed updates and remote attestation protocols that verify integrity before any inference engine starts.

No vendor has yet announced a commercial die that merges CPU, GPU, and tensor cores with unified memory at the 200 GFLOPs level. However, both Intel and AMD have publicly stated that their next‑generation mobile silicon (expected in late 2026) will target

<

70 W TDP while delivering >150 GFLOPs for inference. The key differentiators will be:

• Unified memory architecture eliminating PCIe‑like transfer overhead.

• Higher clock rates for CPU and GPU cores thanks to improved thermal design.

Enterprises should start pilot programs with current silicon families now, while tracking vendor roadmaps. Early adoption of unified architecture will simplify future migration and reduce integration complexity.

• Map Workloads to Silicon Capabilities : Benchmark your most common inference models on i5‑13600H, Ryzen PRO 7000, and RTX 4060 configurations. Use the TechBench AI Suite or equivalent tools.

• Run an Edge‑First Pilot : Deploy 15–20 AI‑Ready laptops in a controlled environment, measuring latency, throughput, battery life, and power consumption against cloud baselines.

• Implement TEE Early : Enable Intel ISE or AMD Secure Enclave on all devices. Validate the ~10 % performance impact and ensure compliance with EU AI Act guidance.

• Negotiate OEM Bundles : Leverage bulk purchasing to secure firmware signing services, extended warranties, and pre‑installed SDKs for ONNX/PyTorch.

• Establish Governance Policies : Define model versioning, audit trails, and access controls at the device level. Align with ISO 27001 and NIST CSF frameworks.

• Plan for Unified Architecture Transition : Monitor Intel’s and AMD’s 2026 silicon roadmap. Allocate budget for future migration once unified dies become available.

In sum, 2025 has delivered a generation of laptops that combine respectable inference performance with power efficiency and emerging security features. For enterprises looking to reduce cloud spend, improve data privacy, and accelerate AI adoption across the workforce, an edge‑first strategy is now technically viable and economically compelling.